Characterizing the Impact of TCP Coexistence in Data Center Networks

Abstract

The switch fabrics of today’s data centers carry traffic controlled by a variety of TCP congestion control algorithms. This leads us to ask: how does the coexistence of multiple variants of TCP on shared switch fabric impacts the performance achieved by different applications in data centers? To answer this question, we conducted an extensive set of experiments with coexisting TCP variants on Leaf-Spine and Fat-Tree switch fabrics. We executed common data center workloads, which include streaming, MapReduce, and storage workloads, using four commonly used TCP variants, namely BBR, DCTCP, CUBIC, and New Reno. We also extensively executed iPerf workloads using these 4 TCP variants to purely study the impact of the coexistence of TCP variants on each other’s performance without incorporating the network behavior of the application layer. Our experiments resulted in a large set of network traces comprised of 160 billion packets (we will release these traces after publication of this work). We present comprehensive observations from these traces that have important implications in ensuring optimal utilization of data center switch fabric and in meeting the network performance needs of application layer workloads.